Anthropic Releases Guide and Reference Implementation for Building Effective Commerce Agents

Anthropic has published a comprehensive guide and reference implementation for building commerce agents that streamline online purchasing and selling. The core architectural recommendation is a single-agent model equipped with agent skills, rather than using intent routers or subagents. This approach preserves shared context across multi-turn conversations, avoiding the state loss, increased token costs, and latency penalties associated with subagent handoffs.

Architecture and Tooling

The guide distinguishes between system prompts and skills based on frequency: content relevant to more than one-third of traffic belongs in the system prompt, while long-tail capabilities are loaded as skills. For UI interactions, the guide advocates treating UI components as tools (e.g., present_products) rather than using custom markup. This allows the server to validate and enrich data, ensures native formatting in the message history, and enables the agent to track screen state for subsequent references.

Latency and Cost Optimization

To minimize end-to-end latency, the guide suggests three levers: fewer turns, faster tools, and faster tokens. Strategies include loading likely context upfront, using parallel tool calls, and dispatching tools eagerly as arguments stream. For perceived latency, streaming components as they form and displaying progress lines during context gathering are recommended. Prompt caching is identified as the largest cost reduction opportunity, with cached input tokens costing 1/10th of fresh tokens. Optimal deployments aim for a 90–99% cache hit rate by structuring requests into Global, Session, and Volatile segments.

Production Operations

In production, memory should be stored externally and written asynchronously to avoid impacting conversation latency, with a three-tier reading strategy (always in context, pre-fetched, and lookup tool). Safety enforcement must occur in the harness, not the prompt; the model proposes actions, but humans or policy apply them via server-issued IDs and strict validation. Evals should focus on final state and rendered responses rather than agent paths, using snapshot-based testing and simulated-user checks for coverage gaps. The guide emphasizes collaboration with SMEs to build 50–100 eval cases per user flow, covering core requests, context dependencies, safety, and multi-capability scenarios.